Downloaded 43 times
![b+tree 数据结构
15| |56| | 77
15 | |30| | 45 … …
22 row
20 row
17 row
16 row
[16, 17, 20, 22]
31 row
35 row
44 row
32 row
[31, 32, 35, 44]
page
有序数组](https://image.slidesharecdn.com/aa-170102081923/75/mysql-acid-4-2048.jpg)
![[OpenInfra Days Korea 2018] (Track 3) - CephFS with OpenStack Manila based on...](https://cdn.slidesharecdn.com/ss_thumbnails/36openinfra-2018-naver-cephfs-180704055415-thumbnail.jpg?width=640&height=640&fit=bounds)
分析mysql acid 设计实现
- 1.
- 2. innodb 基本结构 innodb acid概念 innodb buffer pool mvcc 各种延伸 锁, 各种锁 List
- 3. innodb的抽象结构 tablespace leaf node segment no-leafsegment rollback segment … extent … page … row row row row row row trx id rollback ptr field ptr … segment extent page row
- 4. b+tree 数据结构 15| |56|| 77 15 | |30| | 45 … … 22 row 20 row 17 row 16 row [16, 17, 20, 22] 31 row 35 row 44 row 32 row [31, 32, 35, 44] page 有序数组
- 5. file system vsdisk disk - - - > 512 bytes file system - - -> 4k innodb engine - - - > 16k
- 6. size不同如何上下对齐? partial page write? 解决办法是 doublewrite buffer + doublewrite doublewrite 性能影响? 5% - 10%
- 7. dirty page dirty page doublewrite buffer (2m) double write | double write 数据 文件 redo log crash recovery !!!
- 8.
- 9.
- 10.
- 11. isolation mvcc 隔离级别 READ UNCOMMITTED READ COMMITTED REPEATABLEREAD SERIALIZABLE
- 12.
- 13. innodb buffer pool pagebuffered page buffered page buffered page buffered page buffered page buffered old 37% new 63% 新数据 新数据 新数据 新数据
- 14. update where blog= ‘xiaorui.cc’ data in buffer pool Load page to buffer pool update data and mark dirty page !!! yes ! no !
- 15. innodb buffer pool lru dirtypage free list flush list
- 16.
- 17. if not checkpoint 写操作都是在bufferpool进行，每次提交都会顺序io增加redo日志. buffer pool里面的dirty page越来越多 . redo 越来越大 , 导致recovery时间过长 .
- 18. if checkpoint 当 bufferpool 内存不够用时, 将dirty page刷新到文件. 当redo文件过大又不可循环写入时 , 同上 缩短crash recovery时间
- 19. checkpoint + lsn 待清理Redo Log p1 p2 p3 buffer pool flush list 创建checkpoint lsn: 1500 作为下一个点 p4 p5 刷新到磁盘 ! last checkpoint : 1000 顺序: redo log == flush list ?
- 20. 一个事务过程 1. begin 2. undolog 3. udate where id = xxx; 4. redo buffer 5. redo fsync; binlog; commit if crash ? if crash ? if crash ? one of three happen crash ?
- 21. redo when redo save? when commit ! 顺序io page number + block when redo reduce ? Ringbuffer full flush timer run
- 22.
- 23. flush log rule innodb_flush_log_at_trx_commit= 1 Log Buffer File System Buffer Lb_log file innodb_flush_log_at_trx_commit = 0 innodb_flush_log_at_trx_commit = 2 write os cache timer flush every 1s
- 24.
- 25.
- 26.
- 27. mvcc流程图 trx_id 10 trx_id 15 |row| trx_id | rollback ptr | … read view list trx_id roll ptr data undo data trx_id roll ptr data … undo dataread & write trx_id in read view <= trx_id
- 28. 幻读 record lock gap lock nextgap lock
- 29. 业务数据不一致 事务 A 事务B tsbegin ts select stock where ticket=10 ts begin ts update stock=stock-1 ts commit ts select stock where ticket = 10 ts update stock=stock-1 ts commit stock=1 stock=0 stock=-1 最后 -1
- 30. 互斥锁 vs 乐观锁 互斥锁 select* from train where ticket = 10 for update; 乐观锁 while 1: update train set stock=stock-1 where ticket = 10 and stock = x;
- 31.
- 32. “ Q &A” –fengyun rui